“Generalization in language models, as seen with GPT-2, began to emerge only when models were trained on a broad internet scrape from sources like Common Crawl or Reddit links, rather than narrow datasets.”